01 · AI intake

I scaled myself: four months of intake became one call and a prototype the next week.

Intake was the most manual part of my job. I turned my method into an agentic workflow with three human approvals, which made me faster. Then I made it self-service, so colleagues can start an intake without waiting for me.

4 mo → ~1 wkfrom an AI request to a working prototype

My roleDesigned, built, shipped
WhenAug 2025 – present
Scope23 practices, multiple internal functions, and one global institute
StatusIn the firm's skills catalog, and an agent on our team hub

I run intake for AI requests across the firm: 23 practices, multiple internal functions, and one global institute, each on its own tech stack. I'm not a subject-matter expert in any of them. A request would arrive as a deck, an email, or a Slack message. I'd spend weeks of calls getting up to speed, and I still couldn't get the data to start a proof of concept. Some intakes took four months before anyone touched anything technical. I was running two or three a day, with a lot of context switching, and I couldn't get requirements to my development team fast enough.

I designed and built the agentic workflow myself, then ran it on my own intakes for a month or two until it held up. I took it through peer and security review into the firm's catalog and rolled it out one audience at a time: me, then my team's PMs, then anyone with catalog access, then any colleague with a browser. Then I turned the skill into an agent on our main hub, so people could get to it quickly.

Make the method repeatable, then keep the judgment human. Fixed steps run the same intake every time, use a fraction of the tokens an open-ended chat burns, and pause for approval before anything is written. That gave the team consistency without pretending the system could make the final call.

The approval is the easy part. Deciding what to send through it, and how much reviewer attention to spend, is the design.My working rule
  1. record the callone recorded intake, plus screenshots
  2. extract + score20+ fields, each with a confidence mark; desirability, feasibility, viability
  3. three yeseswho was in the room, whether the summary is right, whether the stories get written
  4. ticketsJira-ready stories, with a first technical step for the engineer

MCP servers for Jira, Slack, email, and the browser add the context so nobody pastes it in. Nothing is created until the third yes.

INTAKE-142Storywritten by the intake skill · made-up example

Sort incoming supplier contracts by risk, so analysts read the risky ones first

Requested by
A procurement team
sure
The problem
Two analysts read every contract by hand to find risky clauses; a batch of 200 takes about two days. We keep missing the auto-renewals.
sure
Who it's for
Procurement analysts; legal reviews anything flagged
sure
Data
About 1,200 past contracts with risk labels; access not confirmed
check
Wanted · buildable · worth it
4 · 3 · 4, out of 5
sure
What good looks like
Flags at least 9 of 10 clauses an analyst would, and never approves a contract on its own
sure
First technical step
Pull 50 labeled contracts and test clause extraction against the analysts' picks
sure
Source: recorded working session, 38 min; every field links to the quote it came fromspeakerssummarystories
A made-up ticket in the shape the intake skill writes. Each field carries a confidence mark, the one it isn't sure about is flagged for a person to check, and nothing is filed until three yeses.
Three ways into one intake workflow: an agent on the team hub, the skill in the firm's catalog, and a web page for people without an AI tool. All three feed the same steps: record the call, extract and score, three yeses from a person, then stories in Jira. an agent on our team hubthe skill, in the firm's cataloga web page, no AI tool needed record the callextract + scorethree yesesstories in Jira
Three doors, one workflow: whichever way someone comes in, the same steps run, and a person says yes three times before anything is filed.
Built withAgentic skillMCP servers: Jira, Slack, email, browserConfidence scoringDFV + RICEWeb front end
4 mo → ~1 wk
from a request to a working prototypebaseline on a recorded call; the same records let us analyze the types of requests teams worked on, to show team impact
1 in 3
of this year's intake requests came in through the skill
$0.3–1.4M
a year of product and engineering time, at 60–100 intakes15–32 product hours and 8–16 engineering hours saved per intake, $200–300/hr; assumptions stated, not measured

People chose it. About a third of this year's intake (32.9%) came in through the skill, and teammates now file their own stories with it. One intake call became a proof of concept the next week, which the team took to its sponsors to secure funding. Every extracted field carries a confidence mark, so the person approving can check it instead of trusting it, and the skill passed peer and security review before it went into the catalog.

I scaled myself. The most manual part of my job is now a self-service front door. Anyone at the firm can start an intake three ways: the agent on our team hub, the skill in the firm's catalog, or a web page for colleagues who don't have an AI tool. Requests also arrive in a form that can be compared, which made the AI portfolio strategy analysis possible.

  • Encode the method you already run. The skill didn't invent a process. It made mine repeatable by people who aren't me.
  • Three audiences, one source. Advisors, PMs, and engineers each need a different output. Design those before the prompt.